LOne A Novel Approach Towards Fake News Detection Using Customized Bidirectional LSTM.pptx
- 1. Fake News Detection Using Custom LOne Model Ankit Shaw (22MAI0046) Samim Aktar (22MAI0047) Romak Das (22MAI0056) Guided by: Dr. Goutam Majumder
- 2. FAKE News is spreading like Wildfire. This needs to STOP !!!
- 3. Introduction • Fake news played a significant role in shaping the 2016 US presidential election. • In developing countries like ours, social media platforms are a nest of fake news where celebrities and sportsmen are mere pawns at the hands of the news writers. • Fake news causes clash among religious communities. • Fake reviews in e-commerce sites harms the brand image.
- 5. Datasets Used Kaggle Fake News Fake Samples Real Samples 10387 10413 Fake Samples Real Samples 37106 35028 WELFake Dataset Widely used dataset in NLP and Machine Learning
- 7. The proposed model • One hot encoding – vectors of length 50 have been used with a vocabulary size of 30000. • Vectorized model passed into a Bidirectional Layer of LSTMs with 100 memory cells. • Dropout layer with a rate of 0.1 is added. • Dense Layer containing 1 node because classification is binary and a value 1 represents Fake News.
- 8. Results Confusion Matrix for WELFake Dataset Confusion Matrix for Kaggle Fake News Dataset
- 9. Comparisons of LOne with existing models Algorithm/Model Accuracy Naïve Bayes 92.12 SVM 96.73 LOne 99.37 Algorithm/Model Accuracy Opt-Net 97.86 WELFake 92.60 LOne 99.37
- 10. Discussions Why use One-Hot Encoding? Each binary feature in one-hot encoding is independent of the others. This means that changes in one category do not affect the representation of other categories. It avoids any implicit assumptions of ordinality or relative importance. Achieved a much higher accuracy with this method so did not change it.
- 11. Discussions (contd.) Why use Bidirectional LSTMS? Capture Long-Term Dependencies – Allows the model to capture dependencies not only from past information but also from future information, enhancing the understanding of the context and improving the representation of the sequence. Improved context understanding – By considering both past and future context, Bidirectional LSTMs have a better understanding of the overall sequence. Robust to Noisy output – LSTM compensates for missing information by understanding the full context and information of the sequence.
- 12. Discussions (contd.) Why use a Dropout Layer? By adding a Dropout layer, the neural network forces itself to learn more robust representations and avoids over-reliance on specific features or connections. This technique helps prevent the model from memorizing noise or irrelevant patterns in the training data, leading to better generalization and improved performance on unseen data during the inference phase.
- 13. Performance of different activation functions on the model. Activation function Accuracy Sigmoid 99.5% ReLu 78.10% Use of ReLu causes the loss of information in out situation, thus preventing the model to capture patterns within the data. The sigmoid function outputs values between 0 and 1. This makes it well-suited for binary classification tasks where the goal is to predict probabilities or make decisions based on a threshold.. The smoothness of the sigmoid function enables more stable and continuous learning during backpropagation. Thus it is evident that our model greatly relies on gradient descent and backpropagation.
- 14. Conclusions • On applying the model over the testing data, it was evident that news articles published by The New York Times was a very reliable source of information. • Our proposed model can also be deployed into many applications such as detecting fake reviews of online products, leading to consumer ambiguity. • Also, websites may be hosted to check online facts for false information, similar to altnews.in and thequint.com. • Social media platforms could also use this model to check integrity of the content circulated through their feed.
- 15. Future Work • The model can further be extended to work on pictorial datasets including pictures, videos or gifs by using image processing techniques such as CNNs, ResNet or Transfer Learning. • Even audio files can be classified by using RNNs to detect any anomalies within an audio sample.
- 16. Thank You. Background image by rawpixel.com on Freepik